Which redaction style keeps the AI useful?

Once you have found the sensitive parts of a message, you have to decide what to put in their place. PromptScrub offers three options, and the choice affects the quality of the answer you get back.

Here is one sentence run through all three:

Email [email protected] and [email protected] again, plus [email protected]. Card 4242424242424242.

Placeholders

Email [EMAIL_1] and [EMAIL_1] again, plus [EMAIL_2]. Card [CARD_1].

Each distinct value gets a numbered token, and the same value always gets the same number. Note that both instances of [email protected] became [EMAIL_1], while [email protected] became [EMAIL_2].

That consistency is the point. The model can still tell that two of the three mentions are the same person, which means it can write "reply to the first address" or "these two are duplicates" and be correct. The sentence keeps its grammar and its logical structure; only the values are gone.

This is the default because it is the only one of the three that costs nothing in answer quality.

Partial mask

Email j***@acme.com and j***@acme.com again, plus b***@acme.com. Card **** **** **** 4242.

Enough survives to recognise the value without exposing it. The masking is tailored per type: emails keep the first character and the whole domain, cards keep the last four digits, in line with how card numbers are conventionally displayed on receipts.

Useful when you need to talk about which item is which and the placeholders would be too abstract, or when the domain itself is part of the question. The cost is real though: acme.com is still on screen, and for many purposes the domain is the sensitive part.

Remove

Email  and  again, plus . Card .

Safest, and clearly the most damaging to the sentence. What is left is barely English. A model asked to work with this has to guess what was removed and how many distinct things there were.

The right use is text where structure does not matter: a paragraph you want summarised for tone, a document where the entities are irrelevant to the question. If you are asking the model to reason about relationships between the redacted things, removal takes away exactly what it needs.

Numbering across a longer text

The consistency guarantee holds for the whole document, not just nearby mentions. Five addresses, three of them identical:

[email protected] [email protected] [email protected] [email protected] [email protected]
→ [EMAIL_1] [EMAIL_2] [EMAIL_1] [EMAIL_3] [EMAIL_1]

Three distinct values, three numbers, and the repeats line up. In a long support transcript this is what lets the model answer "how many different customers are in this thread" correctly.

A note on what placeholders do not hide

Placeholders remove the value and keep the fact. [EMAIL_1] tells a reader that an email address was there, and [EMAIL_1] … [EMAIL_1] tells them the same address appeared twice.

Usually that is exactly what you want, since it is the information the model needs and the value is what you were protecting. But if the mere presence of a category is itself sensitive, placeholders are the wrong choice and removal is the honest one. That is a judgement about your situation, which is why the tool asks rather than deciding for you.